I'm a first-year PhD student in Electrical Engineering at Columbia University. My research centers on signal processing and machine learning, with a particular focus on speech and audio, i.e., building models that exploit the structure of audio and spectral signals rather than treating them as generic sequence data. As a Research Assistant, I developed a Frequency Convolution Network that used dilated convolutions over the log-frequency axis of a Variable-Q Transform frontend for real-time pitch and voicing estimation in just 17.8K parameters, reaching 98.8% RPA50 on MDB-stem-synth, 93.1% on PTDB-TUG, and 97.5% on MIR-1K, alongside spectral-domain augmentation strategies for noise and pitch invariance.
My Master's thesis (completed December 2025) explored how classical digital signal processing and modern deep learning can inform each other for audio synthesis. I built a Variational Autoencoder audio encoder paired with a Mamba-enhanced Differentiable DSP decoder for waveform-level music synthesis and timbre transfer, and designed a self-supervised, EM-based approach to fundamental frequency and voicing estimation that exploits the co-dependence between the two to avoid bad gradient updates on unvoiced frames. The resulting CQT-based pitch/voicing model runs in about 9K parameters, and the full synthesis pipeline is end-to-end trainable from any reference monophonic audio, with no dependence on external F0 or loudness estimators.
This interest in extracting structure from signals predates my time at Columbia — as a research intern at RRCAT, I applied object detection and lightweight CNNs to Raman spectra for pesticide and material identification, including field-deployable TinyML deployments on microcontroller hardware. Alongside this signal processing focus, I've also worked in computer vision (moving-object segmentation for SLAM, GAN-based 3D face editing) and computational biology (multi-omics disease-gene network construction), which continue to inform how I think about representation learning more broadly.
Research Interests
Signal ProcessingAudio & Music MLGenerative ModelsEmbedded ML / TinyMLMechanistic InterpretabilityComputer VisionComputational Biology
Education
Columbia University, USASep 2026 – Present
PhD in Electrical Engineering
New York, NY
Columbia University, USASep 2024 – Dec 2025
MS in Computer Science · GPA: 4.231
New York, NY
IIIT Bangalore, IndiaAug 2019 – Jul 2024
Integrated M.Tech. (B.Tech. + M.Tech.) in Computer Science and Engineering · GPA: 3.85 / 4.00
Bangalore, India
Experience
Columbia UniversityResearch Assistant · New York, NYJan 2026 – Aug 2026
Designed a Frequency Convolution Network (FCN) using dilated 1D convolutions along the log-frequency axis of a Variable-Q Transform (VQT) frontend, achieving a large harmonic receptive field with only 17.8K parameters for real-time monophonic pitch and voicing estimation.
Developed spectral-domain noise and frequency-scaling augmentation strategies to induce noise invariance and pitch equivariance, evaluated on MDB-stem-synth, PTDB-TUG, and MIR-1K with CHiME-Home as a noise source.
Achieved 98.8% RPA50 on MDB-stem-synth, 93.1% on PTDB-TUG, and 97.5% on MIR-1K in clean conditions while keeping RPA50 above 80% at 0 dB SNR across datasets, outperforming a matched-parameter CNN baseline and remaining competitive with substantially larger pitch-estimation models.
Proposed Counterpoint Informed Neural Networks (CINN), embedding Fux's species counterpoint rules as a differentiable, physics-informed penalty on a MinGRU encoder-decoder that generates a soprano accompaniment against a cantus firmus.
Introduced a joint pitch-attack coupling loss (via Jensen-Shannon divergence between neighboring pitch distributions) and an attack-prediction head to keep rhythms and articulations plausible, preventing collapse into rests or rapid spurious attacks.
Implemented a joint beam search over pitch and attack at inference to keep rhythm aligned with the cantus firmus, benchmarking against a MinGRU baseline and a Transformer variant on a held-out composer, reducing rule-based violations across harmonic, melodic, and rhythmic categories with minimal cross-entropy trade-off.
Columbia UniversityMaster's Thesis Project · New York, NYJan 2025 – Dec 2025
Developed a novel Variational Autoencoder (VAE)-based Audio Encoder with a Mamba-enhanced Differentiable Digital Signal Processing (DDSP) Decoder for direct waveform analysis, enabling music synthesis and timbre transfer.
Explored minimal Recurrent Neural Networks (minLSTMs and minGRUs), which significantly improve sequence modeling performance and efficiency by leveraging fewer parameters.
Developed Bidirectional Minimal GRU (BiMinGRU) to effectively incorporate past and future context in the sequence model, removing the reliance on causal convolutions for multi-level architectural designs.
Implemented a self-supervised Expectation-Maximisation (EM) based learning approach to enhance fundamental frequency (F0) and voicing estimation, capitalizing on the co-dependent relationship between existence of fundamental frequency and voicing to filter unvoiced frames and preventing bad gradient updates.
Engineered a highly parameter-efficient frame-level F0 and voicing extraction model with only ≈9K parameters using Constant Q-Transform (CQT) features, prioritizing lightweight and efficient processing for deployment.
Demonstrated a novel fully end-to-end trainable DDSP-based music synthesis and timbre transfer system that accepts any reference monophonic audio, eliminating dependence on external F0 or loudness models allowing for highly specialized models for end-user use cases.
Accessible and Accelerated Robotics (A²R) Lab, Columbia UniversityResearch Assistant · New York, NYSep 2024 – May 2025
Optimized and implemented MobileNet and PyDNet depth estimation models onto resource-constrained microcontrollers (Crazyflie, ESP32) using a custom efficient framework, achieving real-time, on-device depth perception for autonomous systems.
Deployed YOLOv11 for high-speed, real-time object detection in low-power embedded systems, significantly enhancing autonomous target navigation and obstacle avoidance capabilities.
Developed a custom lightweight framework leveraging Teacher-Student Distillation to compress deep learning models into a low-parameter configuration, ensuring efficient resource utilization and reliable operation across diverse microcontroller hardware.
Complex Resilient Intelligent Systems (CRIS) Lab, Columbia UniversityResearch Assistant · New York, NYSep 2024 – Present
Investigated interpretability of LLMs (GPT-2, Gemma 2B Instruct) by analyzing hidden representations using dictionary learning and sparse autoencoders to uncover monosemantic features.
Implemented feature clamping approach (by Anthropic) to steer model outputs toward specific features, improving controllability and interpretability.
Applied Sparse Autoencoders (SAEs) and BERTopic to 500 books across 20 genres, extracting monosemantic features from dense embeddings at multiple granularities to map semantic structure in decoder-based LLMs.
Uncovered unexpected cross-genre semantic relationships through centroid analysis revealing emergent cross-genre relationships validated by narrative archetype theory.
Leading team to investigate emergent scientific understanding in LLMs by training domain-specific and shared-domain SAEs on physics, chemistry and biology textbooks, analyzing how models represent concepts (e.g., atomic structure) differently across disciplines while identifying shared fundamental features.
Developing dead neuron mitigation techniques and hierarchical architectures in SAEs to control learned feature granularity, reducing reconstruction loss while improving interpretability and sparsity.
Digital Video and Multimedia (DVMM) Lab, Columbia UniversityResearch Assistant · New York, NYSep 2024 – Feb 2025
Developed and implemented a novel Vision Transformer approach for Causal Action Understanding, predicting both the action and its resulting object state changes by analyzing start/end frames and bounding box localizations.
Developed an advanced data curation pipeline integrating LLaMA 3.1 via few-shot prompting to resolve issues like object naming inconsistencies and improve action extraction accuracy from raw Ego4D and YouCook2 narrations.
Resolved a critical temporal misalignment in the Ego4D dataset by developing and implementing a binary search algorithm to efficiently map narrations to corresponding video clips, ensuring data coherence for action understanding.
Established performance benchmarks by training and evaluating multimodal baselines (Clip, BridgeFormer, Stateformer, ACTRON) on the enhanced dataset, comparing their capabilities in action prediction and state change localization.
Adapted the YOLO object detection framework for generalized spectra classification, achieving mAP >0.992 for 90 spectra classes and eliminating the traditional need for extensive data preprocessing.
Introduced a novel SVG-based method for automated generation of diverse, annotated spectral image datasets with varied configurations and backgrounds, overcoming limitations of traditional data creation.
Evaluated the integration of attention mechanisms (CBAM and SE) into the YOLO architecture to further enhance feature extraction and improve classification performance in spectral image analysis.
Conducted perturbation analysis on the trained YOLO model to understand and validate how the network was classifying spectra, providing interpretability for the detection mechanism.
Successfully deployed trained YOLO models on smartphones for rapid, real-time spectra classification, benchmarking YOLOv8m and YOLOv8n performance and demonstrating superior efficiency over traditional 1D CNN approaches.
International Institute of Information Technology BangaloreMaster's Thesis Project · Bangalore, INJan 2024 – Jun 2024
Designed and implemented a Panoptic Segmentation Guided Moving Object Segmentation (MOS) framework to correct false positives in the state-of-the-art (SOTA) RigidMask model, specifically enhancing performance for SLAM systems.
Localized the primary cause of SOTA model failure to anomalies (spikes) in 3D P+P cost maps for static objects near frame edges or under occlusion on the challenging VKITTI dataset.
Integrated a top-down information fusion pipeline combining panoptic segmentation (PanFPN), monocular depth estimation (Metric3D), and dense optical flow (RAFT) as a robust post-processing correction step.
Developed an IoU-based panoptic instance tracking algorithm and used dense optical flow matching to estimate per object motion in depth, achieving significantly higher reliability for objects in the 10m - 30m range compared to SIFT.
Engineered a final correction layer using 3D P+P cost histogram analysis to filter residual false positives caused by occlusions or close-range objects, cases where depth-based motion estimation fails.
Achieved substantial quantitative improvements across the VKITTI dataset, including increasing Background IoU to over 99% in tested scenes (e.g., from 87.67 to 99.41 in Scene 01).
Improved the system's robustness by introducing a Reliability Score metric to penalize false positives in frames with no moving objects, with scores reaching 100.0 in scenes where the baseline had high false positive rates.
Statistics and Machine Learning Laboratory, UBC Okanagan CampusResearch Intern · Kelowna, CAAug 2023 – Dec 2023
Developed and implemented a Finite Gaussian Mixture Model (FGMM) with Linear Spatial Covariance structure to perform unsupervised classification of high-dimensional spatial data (MC-simulated radiation dose grids).
Engineered a highly efficient covariance parameterization with only four free parameters, making the model scalable regardless of the number of pixels (p).
Applied an Expectation-Maximization (EM) algorithm to classify biomedical MC simulation data (e.g., 10 mGy vs. 500 mGy dose levels) based on differences in spatial patterns (covariance) rather than global intensity, confirming that higher dose levels exhibit lower spatial variance.
Evaluated two decay functions (Quadratic and Sigmoid) for the spatial covariance component, finding that the Sigmoid decay better approximates complex spatial patterns (“many bumps” vs. “single bump”) relevant to cell/film radiation simulations.
Engineered and optimized diffusion models with multimodal constraints (text and visual inputs) to gain granular control over image synthesis, accelerating and streamlining the product design creation workflow.
Produced visually cohesive, marketing-ready images by leveraging the constrained generative models, ensuring strict adherence to brand references and style guidelines to drive visual consistency across product assets.
Executed advanced multimodal image generation techniques to enhance control and quality, eliminating manual design iterations and significantly improving the efficiency of creating branded visual assets.
Optimized and deployed an 8-bit quantized CNN onto resource-constrained TinyML hardware (Arduino Nano 33 BLE) to achieve accurate classification of 22 complex spectral classes.
Achieved >99.5% classification accuracy on the quantized model by demonstrating the critical role of data normalization, enabling high-precision spectral analysis for field-deployable analytical instrumentation.
Designed and implemented the complete spectral analysis pipeline on-device (preprocessing, ML classification, and analyte quantification) to ensure real-time, autonomous decision-making without external computational or cloud resources.
Investigated and benchmarked state-of-the-art GAN-based models for audio-video generation and style transfer, focusing on applications like realistic lip synchronization (e.g., Wav2Lip) and semantic face editing (e.g., InterFaceGAN).
Implemented and optimized various published GAN architectures for face manipulation (e.g., InterFaceGAN, and models similar to Only a Matter of Style for age regression/transformation), and performed comprehensive testing and comparative analysis using custom datasets.
Combined traditional computer graphics techniques with deep learning by incorporating the FLAME 3D head model into GAN workflows, which improved how realistic and structurally consistent the face edits looked, especially when moving beyond flat 2D images.
Developed SemUV, a deep learning system that generates and edits 3D human heads by working directly with UV texture maps using StyleGANv2, bypassing the limitations of traditional 2D image manipulation approaches.
Demonstrated improved identity preservation and structural consistency compared to 2D methods through user study validation, successfully editing facial features (age, gender, facial hair) while maintaining realistic appearance across different 3D poses.
Built a flexible and efficient solution that works independently of 3D structure, lighting, and rendering settings, allowing quick integration into existing graphics pipelines for VR, AR, gaming, and visual effects applications.
Developed an innovative, low-cost automated system (Arduino UNO & CNC shield) that enabled high-throughput spectral data acquisition (up to 250 spectra/hour), facilitating the generation of large, labeled datasets for ML training.
Implemented a novel raster scanning technique and vision-based automatic sample focus distance control to significantly improve SERS data reproducibility and leverage dynamic 3D hotspots, enabling detection down to a 5 nM limit.
Created an interactive Principal Component Analysis (PCA) GUI for accessible data exploration, allowing non-experts to dynamically evaluate the impact of normalization methods on spectral data clustering.
Developed a novel SERS approach integrating machine learning (PCA & LDA) to achieve >97% accuracy in identifying complex pesticide mixtures, addressing a gap in current analytical methods.
Conducted ML classification and performance evaluation (SVM, kNN, etc.) across diverse datasets, employing Recursive Feature Elimination (RFE) for optimal feature selection and robust classification using limited spectral ranges.
Developed and implemented two novel phenotype-specific filtering methods (network-based and correlation-based) to rank and select significant genes, effectively reducing the large gene space to an optimal size for exploratory visualization.
Constructed a diseasome (disease-gene network) as an unbalanced bipartite graph by integrating multi-omics data (DNA methylation and mRNA expression) for seven cancer phenotypes from TCGA (The Cancer Genome Atlas) using the Heterogeneous Correlation Network Model (HCNM).
Designed a systematic “filter-funnel-focus” workflow for gene prioritization to ensure clutter-free visualization and focused analysis of key genes (e.g., 100+ genes).
Conducted extensive downstream analysis, including pathway enrichment (WikiPathways, Reactome, and KEGG) and functional enrichment (DAVID), to validate biological relevance and compare the performance of the two filtering methods across cancer types.
Developed RadTrixVis, an interactive web-based visualization tool using RadTrix, a novel hybrid graph layout, to display complex disease-gene networks. The layout places a matrix at the center to dramatically reduce visual clutter and edge crossings when showing relationships between diseases and genes.
Integrated an intuitive GUI with interactive features (e.g., phenotype filtering, gene selection, color customization, node reordering) to facilitate disease-specific and gene-specific analysis.
Web Science Laboratory, IIITBResearch Intern · Bangalore, INApr 2020 – Jun 2020
Engineered and deployed a full-stack, interactive COVID-19 trend simulator using React, Leaflet, and Recharts for frontend and a Python/Flask REST API backend to efficiently generate and visualize predictions.
Validated and refined the core SEIR epidemic model implementation, ensuring the correct integration of complex network diffusion and community interaction theory into the production code for accurate trend simulation.
Executed comprehensive testing and troubleshooting of the parameter system and prediction generation logic, guaranteeing reliable simulation results across 10+ user-controlled variables (e.g., mobility, school closures) for policy analysis.
Projects
ArcCell: RNA-seq Cell Type Annotation ToolComputational Genomics Course ProjectPythonJan 2025 – May 2025
Developed a robust cell type annotation tool for sparse scRNA-seq data using ArcFace-based supervised embeddings
Implemented a preprocessing pipeline with GF-ICF weighting, over- and under-sampling to enhance rare cell type representation
Achieved 96% F1-score on test data and 94% on out-of-sample validation, outperforming CellTypist baseline (62%)
Integrated SHAP-based interpretability for gene attribution, finding biologically relevant marker genes across datasets
Linear Regression Accelerator for FPGAEmbedded Systems Course ProjectCVerilogQuartusJan 2025 – May 2025
Designed a memory-mapped 4-bit linear regression accelerator on Intel FPGA with a pipelined datapath for high-throughput accumulation and closed-form inference
Achieved near-floating-point performance (R² = 0.949) using quantized integer arithmetic in hardware
Wrote a custom C-based Linux device driver and user-space interface to control the accelerator and stream training data via AXI-lite to ARM Cortex-A9
RAG based LLM ChatbotPythonLangChainOllamaNov 2024 – Dec 2024
Developed a Retrieval-Augmented Generation (RAG) chatbot using LangChain and Ollama for enhanced context-aware responses
Implemented vector embedding and efficient document retrieval to improve answer relevance and accuracy
Integrated local LLM deployment with Ollama, reducing latency and ensuring data privacy
Reinforcement Learning based ChatbotPythonPyTorchTensorFlowJan 2023 – May 2023
Design two chatbots (Agent A and Agent B) using policy gradient based training strategies so that they can handle prolonged meaningful natural conversations between them
Implement training techniques to give two separate personas for Agent A and Agent B.
Abacus based Mental Calculation using Neuromorphic ArchitecturePythonNengoAug 2022 – Dec 2022
Implementation of different abacus based arithmetic operations to mimic mental calculations using Neural Engineering Framework.
Oceanic Data VisualizationData Visualization Course ProjectPythonSvelteFlaskD3.jsAug 2022 – Aug 2022
Designed a web application to visualize various features of Indian Ocean and Arabian Sea during the period of 29 Dec 2003 to 29 Dec 2005.
2048 Playing AgentPersonal ProjectPythonPyTorchJupyterJun 2022 – Jun 2022
Designed a reinforcement learning agent to play 2048 using different algorithms DQN, PPO, NEAT and Markov Trees.
Image Captioning ToolVisual Recognition Course ProjectPythonTensorFlowJupyterMay 2022 – May 2022
A CNN-LSTM network was implemented using InceptionNetv3 as feature extractor to label various images provided in the publically available Flickr8K dataset.
Auto Rickshaw DetectionVisual Recognition Course ProjectPythonPyTorchJupyterMar 2022 – Mar 2022
Designed object detection model using YOLOv5 on self curated and labelled set of images of auto rickshaw.
Comprehensive Spectroscopic Data Processing and Visualization ToolPythonReactEChartsMar 2022 – May 2022
Designed a full-stack novel software tool for comprehensive data analysis of batch Raman spectra with features like background correction, peaks analysis, heatmaps, SNV, PCA-SVM analysis, spectral data libraries etc.
Utilised React and Apache ECharts for designing frontend and data visualizations of the analysed data provided by backend built using Python and Flask.
Embedded various machine learning models using Python and scikit-learn for determination of compounds for applications like identifying pesticides.
Visual Programming LanguageProgramming Languages Course ProjectJavaScriptBlocklyJan 2022 – Jan 2022
Web based visual programming language to emulate Turtle using Blockly transpiler for equivalent JS code generation.
Web Scraping ToolPersonal ProjectPythonSeleniumSQLiteJan 2022 – Jan 2022
Program to scrape publication DOI and other details from different publisher websites based on user given keywords and record them in a database for ease of querying for researchers.
NLP based Sentiment AnalysisMachine Learning Course ProjectPythonScikit-LearnJupyterDec 2021 – Dec 2021
Machine Learning model for multiclass classification of a sentence into six different types of sentimentally hurting classes; hosted as Community Kaggle Competition.
Face Recognition based Security SystemHackathon ProjectPyTorchScikit-LearnOpenCVJan 2020 – Jan 2020
Real time image pre-processing and face detection using OpenCV and Python integrated with pretrained Deep Learning model to determine strangers and owners.
Publications
Lightweight Self-Supervised Detection of Fundamental Frequency and Accurate Probability of Voicing in Monophonic MusicMar 2026
Venkat Suprabath Bitra, Homayoon Beigi
15th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2026)
Spatial Covariance Constraints for Gaussian Mixture ModelsJan 2026
Hanzhang Lu, Keiran Malott, Venkat Suprabath Bitra, et al. (full author list in paper)
Submitted to Bioinformatics
YOLO-spectra: A generalized framework for rapid simultaneous detection and classification of Raman spectra in images with mobile devices for enhancing on-site applicationsJan 2026
Venkat Suprabath Bitra, Shweta Verma, B. Tirumala Rao
IEEE Transactions on Computational Biology and Bioinformatics, vol. 22, no. 6, pp. 3550-3556, Nov-Dec 2025
HiT-IP Raman: An innovative high-throughput spectra acquisition system with interactive PCA combined machine learning for instant classification of various analytes and mixturesNov 2025
Venkat Suprabath Bitra, Shweta Verma, B. Tirumala Rao
Microchemical Journal, Volume 219, December 2025, 116054, ISSN 0026-265X
Machine learning driven trace detection of pesticide mixtures using citrate optimized Au nanoparticles based in-expensive efficient micro-drop SERS with portable spectrometerNov 2025
Shweta Verma, Venkat Suprabath Bitra, B. Tirumala Rao
Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, Volume 340, 2025, 126333, ISSN 1386-1425
DiffCraft: A Modular Differentiable Framework for Music Synthesis with Timbre TransferMay 2025
Venkat Suprabath Bitra, Homayoon Beigi
IEEE Conference on Artificial Intelligence (CAI 2025)
SemUV: Deep Learning based semantic manipulation over UV texture map of virtual human headsDec 2024
9th International Conference on Computer Vision & Image Processing (CVIP 2024)
Won IAPR Best Student Paper Award
TinyML-Raman: A Novel IoT Based Field-Deployable Spectra Analysis for Accurate Identification of Pharmaceuticals and Trace Dye-Pesticide Mixtures from Facile SERS methodAug 2024
Venkat Suprabath Bitra, Shweta Verma, B. Tirumala Rao
Machine LearningMath for Machine Learning (MML)MML for Signal RecognitionComputer VisionNatural Language ProcessingReinforcement LearningFew-shot LearningCausal Inference and LearningNeuromorphic Computing
Computer Science
Topics in Computability and LearningData VisualizationSoftware TestingOperating SystemsProgramming LanguagesSoftware EngineeringIntroduction to Automata Theory and ComputabilityDatabase SystemsObject Oriented ProgrammingDesign and Analysis of AlgorithmsData Structures and AlgorithmsProgramming
Electronics and Communication
Signals and SystemsComputer ArchitectureDigital DesignEmbedded SystemsControl SystemsComputer Networks